BORD: bayesian optimum radar detector

Signal Processing 83 3) 1151 116 www.elsevier.com/locate/sigpro BORD: bayesian optimum radar detector Emmanuelle Jay a;b;, Jean Philippe Ovarlez a, David Declercq b, Patrick Duvaut b;c a ONERA/DEMR/TSI, BP7, F93 Châtillon Cedex, France b ENSEA/UCP-ETIS, URA D35, BP 44, F9514 Cergy Pontoise Cedex, France c Globespan Inc. Tech., 1 Schulz Drive, Red Bank, NJ 771, USA Received 4 February 1; received in revised form 13 September Abstract We derive the expression of an optimum non-gaussian radar detector from the non-gaussian spherically invariant random process SIRP) clutter model and a bayesian estimator of the SIRP characteristic density. SIRP modelizes non-gaussian process as a complex Gaussian process whose variance, the so-called texture, is itself a positive random variable r.v.). After performing a bayesian estimation of the texture probability density function PDF) from reference clutter cells we derive the so-called bayesian optimum radar detector BORD) without any knowledge about the clutter statistics. We also derive the asymptotic expression of BORD in law convergence), the so-called asymptotic BORD, as well as its theoretical performance analytical threshold expression). BORD performance curves are shown for an unknown target signal embedded in correlated K-distributed and are compared with those of the optimum K-distributed detector. These results show that BORD reach optimal detector performances.? 3 Elsevier Science B.V. All rights reserved. Keywords: SIRP Model; Radar detection; Bayesian estimation 1. Introduction Coherent radar detection against non-gaussian clutter has gained many interests in the radar community since experimental clutter measurements made by organizations like MIT [17,18,4] have shown to t non-gaussian statistical models. One of the most tractable and elegant non-gaussian models results in the so-called spherically invariant random process SIRP) which states that many non-gaussian random processes are the product of a Gaussian Corresponding author. ONERA Palaiseau-DEMR/TSI, chemin de la Huniere, Oce National d Etudes et de Recherches Aerospatiales, Palaiseau cedex F-91761, France. E-mail address: ejay@noos.fr E. Jay). random process called the speckle) with a nonnegative random variable r.v.) the so-called texture), so that an SIRP is a compound Gaussian process. This model is the base of many results like Gini et al. s works [8] in which the optimum detector in the presence of composite disturbance of known statistics modeled as SIRP is derived. In this paper, a bayesian approach is proposed to determine the probability density function PDF) of the texture from N ref reference clutter cells. We use Bayes rule and a Monte Carlo estimation given a non-informative prior on the texture. This approach exploits the SIRP model particularity to describe non-gaussian processes as compound processes and allows one to derive the expression of the optimum detector called bayesian optimum radar detector 165-1684/3/$ - see front matter? 3 Elsevier Science B.V. All rights reserved. doi:1.116/s165-16843)34-3

115 E. Jay et al. / Signal Processing 83 3) 1151 116 BORD). Henceforth, it is no more necessary to have any knowledge about the clutter statistics and BORD deals directly with the received data. In Sections and 3 we briey recall the formulation of a detection problem and describe how the SIRP model clutter yields to general and particular optimum detector. In Section 4, we explain the bayesian approach used to determine a bayesian estimator to the texture PDF and give the expression of the resulting BORD. In this section the asymptotic expression of BORD the so-called asymptotic BORD) and the analytical performance of the Asymptotic BORD are also derived. Section 5 is devoted to the simulations description to evaluate BORD performance compared with optimum detectors performance). Conclusions and outlooks are given in Section 6.. General relations of detection theory We consider here the basic problem of detecting the presence H 1 ) or absence H ) of a complex signal s in a set of N ref measurements of m-complex vectors y = y I +jy Q corrupted by a sum c of independent additive complex noises noises+clutter). The problem can be described in terms of a statistical hypothesis test: H : y = c; 1) H 1 : y = s + c: ) When present, the target signal s corresponds to a modied version of the perfectly known transmitted signal t and can be rewritten as s = AT)t. A is the target complex amplitude and we suppose to have determined all the others parameters ) which characterize the target Doppler frequency, time delay, etc). In the following, p = T )t. The observed vector y is used to form the likelihood ratio test LRT) y) which is compared with a threshold set to a desired false alarm probability P fa ) value: y)= p yy=h 1 ) H 1? : 3) p y y=h ) H LRT performances follow from the statistics of the data: P fa is the probability of choosing H 1 when the target is absent, and the detection probability P d ) is the probability of choosing H 1 when the target is present, that is, P fa = Py) H ) and P d = Py) H 1 ): 3. Non-Gaussian clutter case: SIRV and optimum radar detector In the case of non-gaussian clutter, detection strategies can be derived if we consider a particular clutter nature, i.e. if an a priori hypothesis is made on the clutter statistics. To model non-gaussian clutter and derive general detector expressions, we use the SIRP representation [1,8,13,19]. 3.1. Description and general expressions A spherically invariant random vector SIRV) is a vector issued from a SIRP which modelizes each element of the clutter vector c as the product of a m-complex Gaussian vector x the speckle) CN; M)) with a positive r.v. the texture), that is c = x. The so-formed vector c is, conditionally to, a complex Gaussian random vector CN; M)) with multivariate PDF pc=). The PDF of the clutter is then pc)= m ) m M exp c M 1 c ) p)d; 4) where is the transpose conjugate operator, and M is the determinant of the matrix M. This general expression leads, for a known p), to multivariate PDFs of non-gaussian random vectors. For example, joint K-distributed PDF is obtained if p) is a Gamma PDF see further). 3.. Optimum SIRV detector Applied to the detection problem, expression 4) gives p C y=h ) and p C y=h 1 )=p C y s=h ) when

E. Jay et al. / Signal Processing 83 3) 1151 116 1153 the target signal s is known. The LRT becomes m exp q 1 y)=)d H 1? ; 5) m exp q y)=)d H where q y)=y M 1 y; 6) q 1 y)=q y s) for a known signal s and =ln). When the target complex amplitude A is unknown, ML estimation of A is performed [8]: Â ML = p M 1 y p M 1 p ; 7) and the detection strategy is given by 5) where now [8] q 1 y)=y M 1 y p M 1 y p M 1 p : 8) With 8), expression 5) is called generalized LRT GLRT). 3.3. Examples: various optimum SIRV detectors All the following optimum SIRV detectors are given for an unknown amplitude target whose value is estimated is the ML sense. So, the quadratic forms q y) and q 1 y) are, respectively, given by 6) and 8). 3.3.1. OKD: optimum K detector In the case of K-distributed clutter size m) with parameters and b, texture PDF is a Gamma; =b ) PDF with expression p)= 1 b ) exp b ) : 9) The PDF of y under H hypothesis is then given by p y y=h )= b+m q y) m)= m M ) +m K m b ) q y) ; 1) where K :) is the modied Bessel function of order and :) is the Gamma function. The value of determines the spikiness of the distribution. Following the same processes with 5) expression of the so-called optimum K-distributed detector OKD) becomes m ) m)= q1 y) K m b q 1 y)) q y) K m b H 1? : 11) q y)) H For m =1,q y)= y where y is the scalar vector we observe) and expression 11) becomes y 1 K 1 b y )? H1 1 ) H b 1 : 1) 3.3.. OLD: optimum Laplace detector This detector can be considered as a particular case of OKD since the K-distribution becomes a Laplace PDF when = 1. The texture PDF is then an exponential PDF with parameter b =, with expression p)= b exp b ) ; 13) and the resulting expression of the OLD becomes m ) 1 m)= q1 y) K 1 m b q 1 y)) q y) K 1 m b H 1? : 14) q y)) H For m =1,q y)= y and we just have K b y )? H1 1 H : 15) 3.3.3. OStD: optimum Student-t detector Student-t; b) marginal PDF is obtained when the texture PDF is an inverse gamma PDF IG; =b ) whose expression is p)= b 1 ) exp b ) : 16) The resulting expression of the OStD is then b ) m+ + q y) H1 b? : 17) + q 1 y) H 3.3.4. OD: optimum detector As for the other examples given from now, the texture PDF of a SIRV the marginal PDF is a PDF with degrees of freedom and parameter b; is

1154 E. Jay et al. / Signal Processing 83 3) 1151 116 the variance) is analytically unknown. It is, however, possible [13] to derive the optimum detector for 6 1, which is m k=1 1)k 1 C k 1 m 1 b ) k 1 D k q 1 y);;b) H 1 m k=1 1)k 1 C k 1? ; m 1 b ) k 1 D k q y);;b;) H 18) where for j =; 1, D k q j y);;b;) and i)q j y)) k e b q jy) m 1 = C k 1 m 1 = m 1)! k 1)!m k)! : 3.3.5. OgRD: optimum generalized Rayleigh detector As said previously, the texture PDF of a generalized Rayleigh SIRV the marginal PDF is a generalized Rayleigh PDF with parameters and ; is the variance) is analytically unknown but the optimum generalized Rayleigh detector is given for 6 by ) 1 m ) q1 y) exp q y) [q y) = q 1 y) = ] m 1 k=1 D k)=k!) q 1 y)=) k m 1 k=1 D k)=k!) q y)=) k where is an threshold notation other than, D k )= j 1 k ) 1) j C j j k i ; j=1 i= and C j k =k)!=j)!k j)!. H 1? H ; 19) 3.3.6. OWD: optimum Weibull detector For a Weibull SIRV, the texture PDF can be written as an integral form depending on the G-Meijer s functions. Its expression can be found in [3,9] and the optimum Weibull detector is given for 6 b 6 by ) m q y) q 1 y) m 1 k=1 D kb)=k!)a b ) k F1 kby) exp ab F1 kb y)) m 1 k=1 D kb)=k!)a b ) k F kby) exp ab F kby)) H 1? H ; ) where a, b and are, respectively, the two parameters of the marginal SIRV Weibull and its variance. In the OWD expression, we also have F kb p = q p y) kb for p =; 1; D k b)= and C j k = j 1 k ) 1) j C j jb k i ; j=1 k)! j)!k j)! : i= 3.3.7. ORD: optimum rice detector The Rice marginal PDF of an SIRV, with parameter and variance, gives rise to the optimum rice detector whose expression is given for 6 1by exp[kq y) q 1 y))] m k=1 1)k Cm 1 k =)k D k q 1 y);;) m k=1 1)k Cm 1 k =)k D k q y);;) H 1? H ; 1) where K = =1 ), and for j =; 1 D k q j y)) = k CkI l k l Kq j y)): l= I :) is the rst kind modied Bessel function of order, and Ck l =k)!=l)!k l)!.

E. Jay et al. / Signal Processing 83 3) 1151 116 1155 3.3.8. OCpGD: optimum clutterplus Gaussian detector When thermal noise is separately accounted in the received data, the general expression 5) has to be changed because the total correlation matrix of the noise plus clutter depends on the texture since we have M y= = 1 Eyy =)=M x + bi; ) where I is the identity matrix. GLRT expression 5) becomes 1= M x + b q1y;) I ) exp )p)d H 1? ; 3) 1= M x + b qy;) I ) exp )p)d H where now q y;)=y M x + bi) 1 y; 4) and q 1 y;) is deduced from q y;) as in the previous case known or unknown target amplitude). This case was introduced by Gini et al. in [8] where the authors consider K-distributed clutter plus thermal noise to derive OKGD optimum K plus Gaussian detector). This detector cannot be derived in a simple analytical form but theoretical performances are given in an integral form. From expressions 5) and 3) we can see that SIRV detectors depends on the texture PDF. In [9,1] the authors used a Pade approximation to estimate the texture PDF and then to derive an adaptive detector. This method depends on the quality of the moment estimation and requires reference clutter cells to estimate the Pade coecients. In the next section we propose to use a bayesian estimator of the texture PDF which comes from the Bayes rule and Monte Carlo integration. Then, BORD expression is derived. 4. Bayesian optimum radar detector 4.1. Bayesian study of the problem As we have said in the previous section, for a known texture PDF p), it is possible to derive the associated detector expression. The idea of a bayesian approach is to determine, from N ref reference clutter cells of size m, R =[r 1 ;:::;r Nref ] T where r i = [r i 1);:::;r i m)] T, a bayesian estimator for p). We write p) as follows: p)= p=r)pr)dr: 5) R m Given r N ref a Monte Carlo estimation of 5) is ˆp Nref )= 1 N ref p=r i ): 6) N ref This estimate is unbiased and as the samples {r i ;i= 1;:::;N ref } are statistically independent, the strong law of large numbers applies and gives lim ˆp N ref ) p): a:s: 7) N ref + To evaluate 6) we have to know the expression of the a posteriori PDF of given the N ref reference cells, that is the expression of p=r i ). Using the Bayes rule, we have p=r i )= pr i=)p) ; 8) pr i ) and as p) is unknown as well as pr i )) we replace it by a prior distribution, called g). This processing step works as an information processor that updates the prior density function g) into the posterior PDF p=r i ). The Bayes rule can be interpreted as a relevant mechanism to provide a rational solution of how to learn from the reference cells about the quantity of interest. By this way, Eq. 6) becomes ˆp Nref )= 1 N ref N ref pr i =)g) ; 9) pr i ) where the normalization constant pr i ) is obtained by integrating the numerator in 6) over g) and is given by pr i )= pr i =)g)d: 3) Replacing ˆp Nref ) in 5), the expression of the so-called Bayesian Optimum Radar Detector can be derived. Other methods are investigated in [7,15] to attain detector adaptivity but only with respect to parameters of p). The proposed method makes the detector adaptive with respect to the unknown p) and it is no

1156 E. Jay et al. / Signal Processing 83 3) 1151 116 more necessary to have knowledge about the clutter statistics. 4.. BORD expression The N ref reference clutter cells [r 1 ;:::;r Nref ] T are supposed to be modeled as SIRV and so we have pr i =)= m ) m M exp Z ) i ; 31) where Z i = r i M 1 r i. If we consider this likelihood function as a function of it is proportional to an inverse gamma PDF, IGm +1; =Z i ) an inverse gamma PDF is the PDF of the inverse of a Gamma-distributed variable). In 8) a prior distribution is required for the reference clutter texture and in this case we could choose an inverse gamma prior, IGa p ;b p ) parameters a p and b p would have to be chosen in order to make the prior as non-restrictive as possible). This prior would be the so-called conjugate prior because the resulting PDF would belong to the same PDF family as the likelihood function see [14,16] for more details). As the clutter statistics is unknown we choose a non-informative prior distribution g) = 1=, called Jerey s prior which is proportional to the square root of the Fischer s information measure and which is also an asymptotical case of the inverse gamma PDF when parameters a p and b p +. With the non-informative prior, the a posteriori PDF of given the N ref reference cells can be derived and 8) becomes p=r i )= m 1 ) m M pr i ) exp Z ) i : 3) The normalization constant pr i ) is computed as follows: pr i )= = = m) m M Z m i pr i =)g)d m 1 ) m M exp Z ) i d ; 33) and 3) becomes p=r i )= Zm i m m) m 1 exp Z ) i : 34) This expression is exactly an inverse gamma PDF with parameters m and =Z i. So, we have ˆp Nref )= m 1 N ref m m)n ref Z m i exp Z ) i ; 35) where we recall that Z i = r i M 1 r i. Replacing 35) in 4) for each observed vector y obs size m) and given the N ref reference clutter vectors r N ref we have to compute the following expression under H j j =; 1) to form the GLRT: A N ref m N ref Zi m = m A N ref m N ref m) m 1 exp W ) i;jy obs ) d where Z i is given previously and A N ref m ) 1 =) m M N ref m m); W i;j y obs )=q j y obs )+Z i : Zi m ; 36) W i;j y obs )) m The so-called BORD expression becomes Nref y obs )= Nref [ Z i q 1y obs )+Z i) ] m Nref [ Z i q y obs )+Z i) ] m H 1? H : 37) Both of the quadratic forms q and q 1 are, respectively, given by 6) and 8). BORD expression depends only on the reference clutter cells which provide all the necessary information about the clutter statistics. That makes itself self-adaptive if the correlation matrix is determined from the reference cells of the clutter. This problem was investigated in [5,7] where the authors use the normalized sample covariance matrix estimator NSCM). Given that M = M r =E)=Err )=Er r), the NSCM estimate of M is given by ˆM = m N ref N ref k=1 r k r k r k r ; 38) k

E. Jay et al. / Signal Processing 83 3) 1151 116 1157 1 CFAR Property of the BORD : Simulation with correlated K-distributed clutter ν =.1 ν=.5 ν= ν= 1-1 P fa 1-1 -3 1 1 1 1 1 3 1 4 Threshold values m=1 Fig. 1. CFARproperty of the BORD with respect to the texture PDF shown for correlated K-distributed environments with parameters =:1,.5,, for m = 1. where r k k =1;:::;N ref ) are the N ref reference clutter cells of size m. 4.3. CFAR property of the BORD We can show by simulations that the BORD is CFARConstant False Alarm Rate) with respect to the texture PDF. It means that whatever the clutter statistics is dually whatever the texture PDF is), BORD probability of false alarm is the same. In other words the BORD law is independent of the clutter statistics. This property is shown in Figs. 1 and where the detection threshold is plotted against dierent P fa values. What ever the clutter statistics is K-distributed with various parameters, Gaussian, Student-t or whose texture PDF is a Weibull PDF), the calculation of the detection threshold for a xed P fa is the same for all. 4.4. Asymptotical result of the BORD: AsBORD BORD expression comes after a Monte Carlo estimation of 5) given N ref reference clutter vector r N ref. Given Z i =r i M 1 r i, which is a positive r.v. with PDF pz), BORD expression can be considered as the Monte Carlo estimation of Z m =q 1 y obs )+Z) m )pz)dz Z m =q y obs )+Z) m )pz)dz : 39) Given that r = x where x is a complex Gaussian vector of size m with covariance matrix M, we have Z = r M 1 r = x M 1 x. The quadratic form Q = x M 1 x is m distributed m = Gm; )). So, Z= is Gm; ) and the PDF of Z is derived by integrating pz=) over the prior g). Finally, we can show that lim N ref + N ref y obs )= ) m q y obs ) : 4) q 1 y obs ) This asymptotical result, called asymptotic BORD AsBORD), coincides with the GLRT given, for example, in [15]. It was obtained after replacing the texture by its two maximum likelihood estimates the one under H and the other under H 1 ) in the optimum detection structure 5), where was considered

1158 E. Jay et al. / Signal Processing 83 3) 1151 116 1 CFAR property of the BORD : Comparison wih theoritical expr. of the AsBORD Gaussian K-dist ν=.1) Sudent-t Weibull Texture Theoritical 1-1 P fa 1-1 -3 1 1 1 1 1 3 1 4 Threshold values m = 1 Fig.. CFARproperty of the BORD with respect to the texture PDF shown for dierent correlated environments: K-distributed clutter with parameter = :1, Gaussian clutter, Student-t clutter and unknown clutter whose texture PDF is a Weibull PDF. m = 1. Comparison with the theoretical expression of the AsBORD threshold. as an unknown deterministic parameter. This was also obtained after dierent calculations in [] and for Gaussian clutter in [11,1]. Moreover, with the NSCM estimate of the correlation matrix this expression is also called adaptive linear quadratic ALQ) in [5,6]. 4.5. CFAR property of the AsBORD Replacing both of the quadratic forms q and q 1 in 4) and after simple modication, the AsBORD expression can be written as follows: p M 1 x obs p M 1 p)x obs M 1 x obs ) H 1? H m 1 ; 41) m where x obs is the Gaussian) speckle of the observed SIRV. The left term of 41) is independent of the texture. Then AsBORD PDF is statistically independent of the texture PDF that makes the AsBORD to be CFAR with respect of the texture PDF. In this case it is possible to derive the AsBORD PDF. 4.6. AsBORD PDF According to 41) the AsBORD expression depends on Gaussian vectors and on a related quadratic form. Extending the Cochran s theorem to complex Gaussian vectors and after simple computation, the AsBORD PDF is a Beta PDF with parameters 1 and m 1 whose expression is [9] pu)=m 1)1 u) m : 4) Then the expression of the threshold value depends only on P fa and m, the size of the observed vector [9]: = P m=1 m) fa : 43) Applying Cochran s theorem implies that the covariance matrix of the observed vector is non-singular. Under this condition AsBORD PDF applies also to the BORD as shown in Fig.. In [9] the AsBORD PDF in the case of real vectors is also derived.

E. Jay et al. / Signal Processing 83 3) 1151 116 1159 1 OKD/BORD/OGD Perf. comp. in Correlated K clutter ν=.1, P fa =1-3 ).9.8.7 P d.6.5 OKD BORD OGD.4.3..1-3 -5 - -15-1 -5 5 1 15 SCR input m = 1 Fig. 3. Performance comparison between OGD, OKD and BORD for an unknown target complex amplitude in correlated K-distributed clutter =:1;P fa =1 3 ; m= 1). 1 OKD/BORD/OGD Perf. comp. in Correlated K clutter ν =.5, P fa =1-3 ).9.8.7 OKD BORD OGD.6 P d.5.4.3..1-3 -5 - -15-1 5 5 1 15 SCR input m = 1 Fig. 4. Performance comparison between OGD, OKD and BORD for an unknown target complex amplitude in correlated K-distributed clutter =:5, P fa =1 3, m = 1).

116 E. Jay et al. / Signal Processing 83 3) 1151 116 1 OKD/BORD/OGD Perf. comp. in Correlated K clutter ν=, P fa =1-3 ).9.8.7 OKD BORD OGD.6 P d.5.4.3..1-3 -5 - -15-1 -5 5 1 15 SCR input m = 1 Fig. 5. Performance comparison between OGD, OKD and BORD for an unknown target complex amplitude in correlated K-distributed clutter =, P fa =1 3, m = 1). 1 OKD/BORD/OGD Perf. comp. in Correlated K clutter ν=, P fa =1-3 ).9.8.7.6 OKD BORD OGD P d.5.4.3..1-3 -5 - -15-1 -5 5 1 15 SCR input m = 1 Fig. 6. Performance comparison between OGD, OKD and BORD for an unknown target complex amplitude in correlated K-distributed clutter =, P fa =1 3, m = 1).

E. Jay et al. / Signal Processing 83 3) 1151 116 1161 1 Detection performance of BORD compared with those of OStD in Student-t clutter.9.8.7.6 OStD BORD P d.5.4.3 ν = 1.3 ν =.3..1-3 -5 - -15-1 -5 5 1 15 Input SCR ; m=1 ; P fa = 1-3 / N ref =5 ; N obs =1 Fig. 7. Performance comparison between OStD and BORD for an unknown target complex amplitude in correlated Student-t clutter =1:3; :3, P fa =1 3, m = 1). 5. Detection performances of BORD The correlation matrix M is considered to be known. M values come from a given Gaussian power spectral density and is an m m Toeplitz matrix whose rst auto-correlation coecient is 1 =:98. We compare BORD performance with those of optimum detectors such OKD, OGD optimum Gaussian detector, optimum for Gaussian clutter) or OStD for an unknown target signal embedded in K-distributed clutter or Student-t clutter. In the case of K-distributed clutter, OKD is optimum and we see that BORD performance reach OKD performance whatever the value of is. Dierent values of the shape parameter are tested, =:1; :5; ;. When + K-PDF tends to a Gaussian PDF which is conrmed in the series of Figs. 3 6. In the case of Student-t clutter shown in Fig. 7) the conclusions are the same: BORD reach optimum performances given by those of the optimum detector OStD. All the curves represent the detection probability P d versus the signal-to-clutter ratio SCR) given for one pulse. As m = 1 pulses are considered, the total SCR is 1 log 1 m) = 1 db more than for one pulse. The detection threshold is previously computed via Monte Carlo simulation for each of the detectors to keep a false alarm rate equal to P fa =1 3. Once the detection threshold is obtained we evaluate performance with N ref =1 samples of reference clutter and for N obs = 5 samples of observed clutter data which corresponds in fact to the number of Monte Carlo trials used for the evaluation of the probability of detection) in which an unknown complex target signal for a dierent SCRis embedded. 6. Conclusions and outlooks The present paper has addressed a bayesian approach to the determination of the clutter statistics if we consider the clutter vector modeled as a SIRV. By this way a bayesian estimator of the texture PDF of the SIRV has been derived from reference clutter cells and the resulting BORD expression depends only on these reference cells. For example, in the case of

116 E. Jay et al. / Signal Processing 83 3) 1151 116 CFARConstant False Alarm Rate) detector th reference clutter cells are the cells adjacent to the cell under test. We derive also the asymptotic expression of BORD, called AsBORD, whose PDF is derived in a closed form. With this latter expression we are able to compute the threshold value to set to verify a desired P fa in the case where the correlation matrix is non-singular. Under this assumption, the theoretical threshold expression applies also to BORD. In this paper, we use Monte Carlo simulation to set the threshold value of BORD. In further work we will compare the performances of BORD and AsBORD for experimental ground clutter data and study the applicability of the theoretical threshold expression with respect to the correlation of the data. References [1] T.J. Barnard, D.D. Weiner, Non-Gaussian clutter modeling with generalized spherically invariant random vectors, IEEE Trans. Signal Process. 44 1) October 1996) 384 39. [] E. Conte, M. Lops, G. Ricci, Asymptotically optimum radar detection in compound-gaussian clutter, IEEE Trans. Aerospace Electron. Systems 31 ) April 1995) 617 65. [3] E. Conte, G. Ricci, Sensitivity study of GLRT detection in compound-gaussian clutter, IEEE Trans. Aerospace Electron. Systems 34 1) January 1998) 38 316. [4] A. Farina, F. Gini, M.V. Greco, L. Verrazani, High resolution sea clutter data: a statistical analysis of recorded live data, IEE Proc. F 144 3) June 1997) 11 13. [5] F. Gini, Sub-optimum coherent radar detection in a mixture of K-distributed and Gaussian clutter, IEE Proc. Radar Sonar Navig. 144 1) February 1997) 39 48. [6] F. Gini, M.V. Greco, A sub-optimum approach to adaptive coherent radar detection in compound-gaussian clutter, IEEE Trans. Aerospace Electron. Systems 35 3) July 1999) 195 114. [7] F. Gini, M.V. Greco, M. Diani, L. Verrazzani, Performance analysis of two adaptive radar detectors against non-gaussian real sea clutter data, IEEE Trans. Aerospace Electron. Systems 36 4) October ) 149 1439. [8] F. Gini, M.V. Greco, A. Farina, P. Lombardo, Optimum and mismatched detection against K-distributed clutter plus Gaussian clutter, IEEE Trans. Aerospace Electron. Systems 34 3) July 1998) 86 876. [9] E. Jay, Detection en environnement non-gaussien, Ph.D. Thesis de l Universite de Cergy-Pontoise, June. [1] E. Jay, J.-P. Ovarlez, D. Declercq, P. Duvaut, PEOD: Pade estimated optimum radar detector, IEEE Radar Conference Atlanta, USA, May, pp. 7 74. [11] V.A. Korado, Optimum detection of signals with random parameters against the background of noise of unknown intensity under conditions of constant false alarm probability, Radio Eng. Electron. Phys. 13 1968) 969 97. [1] B. Picinbono, G. Vezzosi, Detection d un signal certain dans un bruit non stationnaire et non Gaussien, Ann. Telecomm. 5 197) 433 439. [13] M. Rangaswamy, D. Weiner, A. Ozturk, Non-Gaussian random vector identication using spherically invariant random processes, IEEE Trans. Aerospace Electron. Systems 9 1) January 1993) 111 13. [14] C. Robert, Methodes de Monte Carlo par Chaˆnes de Markov, Statistique mathe-matique et probabilite, Economica, Paris, 1996. [15] K.J. Sangston, F. Gini, M.V. Greco, A. Farina, Structures for radar detection in compound Gaussian clutter, IEEE Trans. Aerospace Electron. Systems 35 ) April 1999) 445 458. [16] M.A. Tanner, Tools for Statistical Inference, Springer Series in Statistics, Springer, New York, 1996. [17] K.D. Ward, C.J. Baker, S. Watts, Maritime surveillance radar. Part 1: radar scattering from the ocean surface, IEE Proc. F 137 April 199) 51 6. [18] S. Watts Ed.), Special issue on radar clutter and multipath propagation, IEE Proc. F 138 April 1991). [19] K. Yao, A representation theorem and its application to spherically invariant random processes, IEEE Trans. Inform. Theory 19 ) 1973) 6 68.